Multi-task U-Net with class-conditional Mahalanobis scoring for drone acoustic anomaly detection

초록

With the growing use of drones, early detection of abnormal operating conditions has become increasingly important. However, collecting representative fault examples before deployment is often difficult, making anomaly-label-free anomalous sound detection (ASD) practical for drone health monitoring. This study presents a multi-task U-Net that jointly reconstructs log-Mel spectrograms and predicts normal operating-condition metadata, namely drone type and flight direction. Anomaly scores are then computed as operating-condition-conditional Mahalanobis distance (MD) scores whose statistics are estimated only from normal training embed-dings. Gradient-based losses, input-normalization statistics, and normal-reference estimation use normal training samples. Drone-type and flight-direction metadata supervise the auxiliary objectives and define the conditional reference groups. Labeled evaluation data are used for hyperparameter optimization, learning-rate scheduling, early stopping, and checkpoint selection, whereas the test set is reserved for final reporting. On the ICSV31-AI Challenge dataset, the full configuration achieved a test macro-averaged area under the receiver operating characteristic (ROC) curve (Macro AUC) of 0.9592 +/- 0.0059 over five random seeds. In the matched representation-scoring analysis, the multi-task representation with conditional scoring achieved a Macro AUC of 0.9592 +/- 0.0059, compared with 0.4833 +/- 0.0378 for the single-task representation with global scoring. A reconstruction-plus-classification variant achieved 0.9588 +/- 0.0048, indicating that operating-condition-supervised representation learning was the major empirically supported contributor, while the remaining auxiliary components showed mixed or seed-dependent effects. Predicted metadata reduced the Macro AUC to 0.9095 +/- 0.0087, and an exploratory normal-training-score threshold analysis revealed substantial false-alarm rates. The results support condition-aware representation learning and conditional normal modeling while identifying limitations in meta-data reliability, threshold calibration, and cross-domain generalization.

키워드